MS-DRT: A Multilevel and Multiscale Branch Learning Scheme for Fault Diagnosis of Rotating Machinery

• Published in: IEEE transactions on industrial informatics 2024-02, Vol.20 (2), p.2799-2811
• Main Authors: Xu, Xiaogang, Wang, Ruijun, Fan, Zhixia, Ma, Xu, Zhao, Zeren, Wang, Huijie
• Format: Article
• Language: English
• Subjects: Convolution; Data models; Decision making; Deep learning; Densification; Fans; Fault diagnosis; Feature extraction; Kernel; multilevel decision; multiscale branching structure; Performance enhancement; Pixels; Rotating machinery; Vibrations
• ISSN: 1551-3203; 1941-0050
• DOI: 10.1109/TII.2023.3295426

The state detection of mechanical equipment is one of the important application fields of intelligent diagnosis. In this article, a multiscale dense residual network is developed to solve the problem of inefficient diagnosis of mechanical equipment in complex operating environments. The model improves the performance of the model from three aspects: pixel level, feature level, and decision level. At the pixel level, the introduction of data input frameworks with different granularities reduces the limitation of traditional models that only extract fault features on a single time scale, and enriches the utilization of state information of mechanical equipment. At the feature level, multiscale dense residual units are used to perform layer-by-layer explicit amplification and densification learning of global and local features. Make the model learn the fault information of different levels and depths to the maximum extent. In addition, the residual connection of interval sampling is used to correct the coding abnormal behavior to stabilize the diagnostic performance of the model. At the decision level, the original feature decision-making mode is changed, and the proposed regional mean strategy can efficiently aggregate multiscale features. The expression demands of each feature of model learning are considered. Through the state recognition of different mechanical equipment, it is identified that the model has excellent generalization ability. Finally, the general relationship between spectrum characteristics and filter parameters is explored to obtain a microscopic representation that the deep learning model extracts fault signal features.